According to a study published on ArXiv CS.AI, padding prompts makes vision-language models more robust to image corruption, whereas fine-grained questions increase model fragility.
In tests on Qwen3-VL and LLaVA-OneVision, verbose paraphrasing reduced answer drift variance by 70 to 81 percent on 8B models, offering a simple prompt-level defense against visual distortion.
Microsoft Research published RetroChimera in Nature, an ensemble retrosynthesis model it says outperforms prior models in blind tests, and released its code and weights.
MIT CSAIL, with Google and Northeastern, released InstructMesh, which pairs Microsoft's TRELLIS 3D generator with GPT-4 so users can edit and repair AI-generated designs by highlighting parts to fix.
MIT researchers argue in Philosophical Perspectives that algorithmic monoculture isn't uniformly harmful: its main problem is informational echo chambers that hinder exploration, and ensembling algorithms can mitigate this.
Latent Space published an interview with MIT's Alex Zhang covering Recursive Language Models, AI-written GPU kernels, and agent swarms, saying an RLM-based harness was first to roughly solve ARC-AGI-3 before OpenAI's Astra.
An arXiv paper describes RPG, a framework that has a robot system diagnose its own failures in simulation and rewrite its skills and prompt without retraining the model.
An arXiv paper (ArXiv CS.CL) proposes ZFO, which picks the update direction with a standard first-order optimizer, then uses two extra function evaluations to choose a curvature-aware step size.